News

OpenAI signs four groups for embedded model-safety assessments

OpenAI signed agreements with Apollo Research, METR, Transluce, and Goodfire for embedded safety assessments with varying access, oversight, and publication terms.

D
Nov 16, 2026 · 4 min read

OpenAI signed agreements with Apollo Research, METR, Transluce, and Goodfire for a first round of embedded safety assessments, giving the outside groups varying access to internal models, data, monitoring systems, and staff. The four engagements will examine whether models undermine safeguards, whether alignment evaluations miss failures, and whether interpretability tools can expose problematic behavior.

The program does not create one uniform audit. Each assessor has an agreed scope, access, resources, and safety arrangements, while some access terms remain unsettled. OpenAI said assessors may ask to follow new evidence, but the company may use alternatives such as having its researchers run an analysis when intellectual-property concerns prevent direct access.

What each group will examine

Apollo will test four claims: whether training created incentives for models to subvert safeguards; whether evaluations and red teaming elicit that behavior; whether models attempted it during internal use; and whether related reasoning would be detectable. Apollo says six evaluators will spend four months inside OpenAI across training, evaluation, and internal-deployment work.

For Apollo’s first stage, OpenAI said it will provide selected public and unreleased models and training checkpoints, relevant training data and evaluation results, intervention records, monitoring systems and flagged behaviors, staff access, tools, and compute. Apollo said detailed access for three of the four claims is still being negotiated. OpenAI can withhold in-scope material for technical infeasibility, security risk, or intellectual-property protection, according to Apollo, but must document refusals and cannot withhold material solely because it is unfavorable.

METR’s first alignment assessment will look for misaligned behavior and missed failures in OpenAI’s largest frontier reinforcement-learning run. OpenAI said METR can set its own priorities within the selected run and agreed arrangements, with access to relevant checkpoints, training and evaluation rollouts, and misalignment-monitoring records. The scope is narrower than METR’s published proposal, which describes a wider survey across training, evaluations, and internal use. OpenAI and METR have also preliminarily agreed to a broader monitor assessment, but have not finalized its scope.

Transluce said its initial on-site pilot will begin in early December and run for six weeks. It will assess the scope, quality, and coverage of OpenAI’s post-training alignment evaluations, including whether they capture deception, manipulation, and reward hacking. Under Transluce’s stated terms, the group will get privileged access to a handful of models, including at least one internal-only model and raw chain-of-thought outputs, as well as selected non-public evaluations, prior training and evaluation transcripts, safety-case documentation, on-site infrastructure, and some use of its own tools. Transluce stressed that this is not employee-level access: it may request more material, but OpenAI decides whether to grant it.

The parties intend a roughly eight-week follow-on assessment in early 2027 with broader model access and more information about alignment evaluations and how their results affect release decisions. Transluce said those details are not final.

Goodfire will embed interpretability researchers who use its platform to analyze model internals. OpenAI identified training-run monitoring, debugging, and alignment as the main objectives, with OpenAI researchers as the primary audience and public findings expected as work progresses. Goodfire has separately demonstrated activation probes for detecting reward-hacking signals, but that research does not establish the scope or results of this engagement. OpenAI did not specify Goodfire’s start date, duration, or exact internal access.

Oversight, disclosure, and publication

Assessors handling confidential information will work on site using OpenAI devices and must follow at least the company’s employee-level security standards. OpenAI said its Foundation Board’s Safety and Security Committee will have visibility throughout the work, receive unredacted reports, and be able to speak privately with assessors. Unresolved disputes will move through OpenAI’s Safety Advisory Group and executives before referral to the committee.

Employee conversations will be off the record and cannot be published or attributed. Assessors may instead ask OpenAI to attest to a statement, then publish the attested statement or disclose that the company declined to attest. OpenAI also said participating staff are asked to observe an unspecified brief cooling-off period before embedding at another AI developer to reduce the risk of carrying confidential information between companies.

OpenAI said the assessors remain independent in publication, while it retains the right to redact confidential information and may request a delay to address safety or security findings. The exact limits vary. Apollo plans two public reports covering findings, methods, and limitations; it says OpenAI cannot redact Apollo’s verdicts or seek a redaction merely because a finding is unfavorable. Apollo may also disclose withheld or delayed access, report removals, and perceived pressure without OpenAI pre-review of those disclosures.

Transluce says it has editorial control, but OpenAI may request redaction of intellectual property or other non-public information, and results about internal-only models cannot be published. Its unredacted report, like Apollo’s, will go to the Safety and Security Committee.

Conflict procedures will cover personal relationships, financial interests, recusals, and transparency. Apollo and Goodfire will receive compensation that OpenAI says is not tied to results. METR and Transluce will not accept payment, though they may accept API credits or compute. OpenAI said the protocols are not standardized and are expected to change as the first assessments expose practical issues involving technical access, security, and workspace.

More news